SeedVR2

Video restoration with diffusion used to mean dozens of denoising passes per frame window; SeedVR2 collapses that to a single step through adversarial post-training, making diffusion-quality upscaling practical on real footage. The ByteDance Seed models are diffusion transformers with adaptive window attention, so they handle arbitrary resolutions and lengths without tiling artifacts, sharpening compressed, blurry, or AI-generated clips and working on still images too. Weights come in 3B and 7B sizes under Apache-2.0 alongside the earlier multi-step SeedVR, a CVPR 2025 highlight; SeedVR2 itself was accepted to ICLR 2026. Official scripts target big iron, running 720p on a single 80GB H100 and 1080p or 2K across four of them, but community ComfyUI nodes squeeze the 3B model onto consumer cards, and that path has become the standard finishing stage for AI video pipelines. The repo ships inference code and one-command demos with weights pulled automatically from Hugging Face, and the maintainers document honest failure modes like over-generated detail on lightly degraded input.